Skip to content
Hands-on IT IT Drills

Day 0: Training Ground

Goal: A real three-node cluster, a victim and an attacker pod are ready. · Duration: approx. 60 min · Prerequisite: three VMs

Briefing

A toy cluster on your laptop is not enough here: many attacks target the node – kubelet, node file system, etcd. For that you need real machines. We use RKE2 on three VMs: one server, two agents.

  rke2-server .126         rke2-agent1 .228      rke2-agent2 .83 
  control-plane + etcd     worker                worker
        |________________________|___________________|
                     one cluster, three nodes

On top of that, two roles in the cluster: the namespace shop is the “normal” application (the victim), the namespace attacker is the location of a compromised workload. From there you run the attacks – never from your laptop with cluster-admin, but the way an attacker would sit: in a pod, with its permissions.

Drill 1: Prepare the VMs

On all three VMs (via SSH). Swap off, firewall off in the lab, hostnames set:

sudo swapoff -a
sudo hostnamectl set-hostname rke2-server   # or rke2-agent1 / rke2-agent2
cat /etc/os-release | grep PRETTY_NAME

Expected:

PRETTY_NAME="Ubuntu 24.04.5 LTS"

Save your progress? Sign up for free

Drill 2: Install the RKE2 server

On rke2-server (.126):

curl -sfL https://get.rke2.io | sudo sh -
sudo systemctl enable --now rke2-server.service

Wait until the service is running, and make kubectl usable:

sudo systemctl is-active rke2-server
export KUBECONFIG=/etc/rancher/rke2/rke2.yaml
export PATH=$PATH:/var/lib/rancher/rke2/bin
sudo -E kubectl get nodes

Expected (shortened, may differ depending on the version):

active
NAME          STATUS   ROLES                 AGE   VERSION
rke2-server   Ready    control-plane,etcd    90s   v1.36.6+rke2r1

You will need the join token in a moment:

sudo cat /var/lib/rancher/rke2/server/node-token

Save your progress? Sign up for free

Drill 3: Attach the two agents

On rke2-agent1 (.228) and rke2-agent2 (.83). <TOKEN> is the value from Drill 2, <SERVER-IP> is 172.16.0.126:

curl -sfL https://get.rke2.io | sudo INSTALL_RKE2_TYPE="agent" sh -
sudo mkdir -p /etc/rancher/rke2
cat <<EOF | sudo tee /etc/rancher/rke2/config.yaml
server: https://172.16.0.126:9345
token: <TOKEN>
EOF
sudo systemctl enable --now rke2-agent.service

Back on the server, check:

sudo -E kubectl get nodes

Expected (shortened):

NAME          STATUS   ROLES                 AGE   VERSION
rke2-agent1   Ready    <none>                88s   v1.36.6+rke2r1
rke2-agent2   Ready    <none>                69s   v1.36.6+rke2r1
rke2-server   Ready    control-plane,etcd    5m    v1.36.6+rke2r1

Tip: Put export KUBECONFIG=/etc/rancher/rke2/rke2.yaml and the PATH into the server’s ~/.bashrc; that does not save you the sudo -E – but for the node attacks later you need root anyway. In this book, work as ubuntu with sudo where needed.

Save your progress? Sign up for free

Drill 4: Set up the victim

A plain application in the namespace shop: a web frontend, an API, a database. Still without any hardening – that is intentional.

sudo -E kubectl create namespace shop
sudo -E kubectl -n shop create deployment frontend \
  --image=nginxinc/nginx-unprivileged:stable
sudo -E kubectl -n shop create deployment api \
  --image=ealen/echo-server:latest
sudo -E kubectl -n shop create deployment db \
  --image=postgres:16-alpine
sudo -E kubectl -n shop set env deployment/db POSTGRES_PASSWORD=lab-password
sudo -E kubectl -n shop get pods

Expected (shortened):

NAME                        READY   STATUS    RESTARTS   AGE
api-xxxxxxxxxx-xxxxx        1/1     Running   0          20s
db-xxxxxxxxxx-xxxxx         1/1     Running   0          20s
frontend-xxxxxxxxx-xxxxx    1/1     Running   0          20s

Save your progress? Sign up for free

Drill 5: Set up the attacker pod

The compromised workload. A perfectly normal pod with network tools and its default ServiceAccount token – an attacker needs nothing more as a starting point:

sudo -E kubectl create namespace attacker
sudo -E kubectl -n attacker run shell --image=nicolaka/netshoot \
  --command -- sleep infinity
sudo -E kubectl -n attacker wait --for=condition=Ready pod/shell --timeout=60s

This is how you get into the attacker pod – from now on you are “in the cluster”, no longer at the control desk:

sudo -E kubectl -n attacker exec -it shell -- bash

First look from the inside: which token does the pod carry?

ls /var/run/secrets/kubernetes.io/serviceaccount/
cat /var/run/secrets/kubernetes.io/serviceaccount/namespace

Expected:

ca.crt  namespace   token
attacker

The token in that directory is your key to the Kubernetes API. What you are allowed to do with it is the question of the next days – and how you restrict it.

Save your progress? Sign up for free

Checkpoint

Three nodes ready:

sudo -E kubectl get nodes --no-headers | grep -c Ready

Expected: 3

Victim and attacker are running:

sudo -E kubectl get pod -n attacker shell -o jsonpath='{.status.phase}'; echo

Expected: Running

The attacker pod has an API token:

sudo -E kubectl -n attacker exec shell -- \
  test -s /var/run/secrets/kubernetes.io/serviceaccount/token && echo hat-token

Expected: hat-token

Challenge

Find out which ServiceAccount the attacker pod runs as – without going into the pod, only with kubectl from the server.

Hint: A pod without a serviceAccountName set gets a particular default.

Solution

sudo -E kubectl -n attacker get pod shell \
  -o jsonpath='{.spec.serviceAccountName}'; echo

Expected: default

Every namespace has a default ServiceAccount, and every pod without its own setting uses it. In a well-hardened cluster this default is allowed to do almost nothing – we check that on Day 2.

Save your progress? Sign up for free

Enterprise Note

  • In production, RKE2 comes with CIS hardening (profile: cis) and an admission setup; we deliberately start unhardened to see the gaps.
  • Keep lab clusters strictly separate from the company network. A deliberately vulnerable cluster is a worthwhile target – for others, too.
  • etcd sits on the server node here. In real setups it belongs encrypted and on dedicated nodes with restricted access (Day 10).